Night-time vehicle classification with an embedded vision system

The paper presents night-time vehicle classification using an embedded vision system based on an optical transient sensor. This neuromorphic sensor features an array of 128x128 pixels that respond to relative light intensity changes with low latency and high dynamic range. The proposed algorithm exploits the temporal resolution and sparse representation of the data, delivered by the sensor in the data-driven Address-Event Representation (AER) format, to efficiently implement a robust classification of vehicles into two classes, car-like and truck-like, during night-time operation. The classification is based on the extraction of the positions and distances of the vehicles head lights to estimate vehicle width. We present the algorithm, test data and an evaluation of the classification accuracy by comparison of the test data with ground truth from video annotation and reference results from a state-of-the-art ultrasonic/radar-combination reference detector. The results show that the difference in total truck counts with respect to a reference detector and to manually annotated video during night-time operation under dry and wet road conditions is typically below 6%.